Reading the Leader's Push: Calibrated Haptic Evidence for Occlusion-Aware Co-Carrying

When a person and a robot carry a table, stretcher or box together, the load usually blinds the robot. The person in front can see the corridor, the doorway and the child running around the corner. The robot behind cannot. Today's systems handle this mismatch in one of two opposite ways. They either follow the human's pushes and trust the human completely, or they ignore the human and plan against a worst-case guess about everything hidden. The first is unsafe whenever the human misses something. The second makes the robot brake at every blind corner and pull back against a partner who is walking forward.

We propose a middle path. The robot treats the force the leader applies as evidence about what the leader sees. It turns that evidence into a calibrated set of plausible next maneuvers drawn from a small library of discrete modes such as "continue", "slow", "stop", "turn left" or "rotate in place". It relaxes its worst-case assumptions about hidden obstacles only in regions the human can actually see, and only when every mode in the set agrees that proceeding is safe. In two sentences: the leader's push is a sensor reading about the robot's blind spot, and with the right statistics that reading can safely shrink the robot's paranoia. We make the reading trustworthy by predicting over a discrete set of maneuvers, which supports cheap, distribution-free coverage guarantees, and not over a continuous velocity.

Core Claim

A haptic intent estimator that outputs conformal prediction sets over discrete co-carrying modes can serve as a certified gate on an occlusion-aware safety filter. A conformal prediction set is a set of labels that contains the true label with a user-chosen probability, $1-\alpha$, under only an exchangeability assumption. The gate can remove a large share of the robot's resistance to the human, measured as linear impulse, and of its deadlocks at blind corners. It adds collision risk only in a quantifiable way: roughly $\alpha$ plus the rate at which the leader fails to perceive hazards inside their own field of view. The system never relaxes safety in regions neither partner can see.

The idea transplants the "other agents are sensors" reasoning of Behavior Planning of Autonomous Cars with Social Perception. That work infers a hidden pedestrian from a nearby human driver's braking by assuming the driver is noisily optimal. We move the reasoning from remote visual observation of other cars to direct physical contact with a partner. We also replace its rationality assumption with an empirical coverage guarantee, and we confine the trust to the partner's estimated visible region.

What Exists and Where It Stops

Three lines of work each supply one ingredient and stop short of the combination.

The first line shows that pushes carry intent. Bajcsy et al. and Losey et al. treat physical corrections as communication about the human's objective. Constraint-Aware Intent Estimation for Dynamic Human-Robot Object Co-Manipulation estimates continuous intent from velocity tracking errors with particle filters. It prunes kinematically infeasible hypotheses and cuts human linear impulse to 76% of standard admittance control. These papers motivate the proposal, but all of them work in open, fully visible spaces.

H2-COMPACT is the closest existing work. A Unitree G1 humanoid co-carries a box whose bulk blocks its cameras, so it infers velocity purely from wrist force sensing through a diffusion policy. It explicitly delegates all collision avoidance to the human. It has no model of hidden hazards and no mechanism for doubting the leader.

The second line supplies safety under occlusion. OA-MPC grows reachable "capsule sets" from occlusion boundaries at an assumed maximum hidden-agent speed. It guarantees collision avoidance, but the file notes that this worst-case bound drives freezing in cluttered corridors. We reuse its reachability machinery as our safety filter and attack its conservatism. Collaborative perception such as V2VNet and CoBEVT removes occlusion by sharing sensor features between agents. It needs infrastructure and a network, neither of which exists when two people carry a sofa through a stairwell.

The third line explains why the discrete choice matters. RCIP builds conformal intent sets and maps them to discrete actions: it acts when all plausible intents imply the same action and asks the human otherwise. It works only on semantic tasks with no contact and no occlusion. We borrow its "act only when the set agrees" logic directly. Tac-Nav shows that discrete directional intents can be classified from handle pressure with 90.8% real-world accuracy, which supports a discrete mode vocabulary as a workable target.

The literature also constrains the design. Discrete mode switches produce jumps in velocity and acceleration that destabilize compliant control, as argued by Neural Implicit Action Fields and noted for diffusion policies by Duan et al.. Our architecture therefore keeps the discrete layer strictly as a decision and certification layer. Execution stays continuous and compliant.

Research Program

The system has three parts.

Haptic mode predictor. This component maps a short window of wrist force and torque plus velocity tracking error, $f_{t-k:t}$, to a score over a mode library $\mathcal{M}$. We seed $\mathcal{M}$ with the eight motion primitives already labeled in the H2-COMPACT human-human dataset, and add "slow" and "stop". The label we predict is what the leader does next, not what the world contains. This distinction is the honest core of the method, because the guarantee concerns predicting the human's behavior. We calibrate on held-out trials with split conformal prediction, which picks the score threshold from held-out data so the coverage guarantee holds without distributional assumptions:

$$C_\alpha(f_{t-k:t}) = \{\, m \in \mathcal{M} : s(f_{t-k:t}, m) \le \hat q_{1-\alpha} \,\},$$

Here $s$ is the nonconformity score, meaning one minus the predicted probability of mode $m$. The threshold $\hat q_{1-\alpha}$ is the calibration quantile, so that $P(m^\star \in C_\alpha) \ge 1-\alpha$ for the leader's true next mode $m^\star$.

Visibility-gated safety filter. We run OA-MPC's reachable sets as usual. The robot estimates the leader's pose from the rigid payload's kinematics, since the leader holds the other end, and from that it estimates the leader's visible region $V_t$. Inside $V_t$, and only there, reachable sets from occlusion boundaries are removed when every mode in $C_\alpha$ implies forward progress. If the set contains "stop" or "slow", or disagrees on direction, the full worst-case sets stay in force. Outside $V_t$, for example the lateral blind corner the leader has already walked past but the robot has not, worst-case sets always apply.

This gives the target risk decomposition. For hazards inside $V_t$, the probability of an unsafe relaxation is at most

$$\alpha + \rho,$$

where $\rho$ is the rate at which the leader fails to react to a hazard they could see. The $\alpha$ term follows from conformal coverage. The $\rho$ term is irreducible because the method trusts human perception and does not replace it. Measuring $\rho$ is part of the study.

Continuous compliant execution. A variable-impedance controller tracks a reference that the filter's mode decisions modulate through smoothly blended velocity limits, not hard switches. We will test whether an analytic continuous reference in the style of NIAF, compared against simple blending, measurably reduces jerk and force spikes at mode transitions. We will not assume it helps.

Baselines

Pure haptic following. The naive reference point is a well-tuned admittance controller with no occlusion reasoning, standing in for "just follow the human". Alongside it we run the strongest learned follower, a force-driven policy in the H2-COMPACT style. That paper reports a mean trajectory deviation of 0.1294 m against 0.1109 m for human-human pairs, and 16.23 N average leader force against 17.35 N with a human follower. This family is the strongest competitor on effort and fluency, because it never resists the leader. We beat it if we match its impulse within a pre-registered margin while cutting collisions with hazards outside the leader's view. By construction, those are hazards this baseline cannot handle.

OA-MPC with admittance. The strongest safety baseline pairs OA-MPC unchanged with the same compliant controller. For this we found no reported numeric success or impulse result on Lacuna; the paper reports only that it avoided collisions where a baseline MPC did not, at 10 Hz. It is the strongest competitor on safety. We beat it if we keep its zero-collision record on hidden agents outside $V_t$ while lowering linear impulse, completion time and deadlock rate. For reference, Shao et al. report impulse falling to 76% of admittance, a gain we expect to be partly lost under a worst-case filter.

Heuristic trust. This baseline feeds a continuous point-estimate intent, the Shao-style particle filter, into OA-MPC. A single tuned scalar shrinks the assumed hidden-agent speed as intent confidence rises.

Two ablations isolating the key ingredients. The first keeps our full pipeline but gates on the argmax mode instead of the conformal set. This isolates calibration. The second removes the visibility gate $V_t$ and relaxes everywhere, which isolates the "trust only what the leader can see" ingredient.

All methods share the same simulator episodes, the same human-behavior models, the same force signal and the same low-level controller. Each gets an equal hyperparameter search budget, on the same number of tuning episodes disjoint from the test episodes.

Why not just…?

Why not just follow the human, since the human can see? This is the real threat. In scenarios where every hazard lies in the leader's view and the leader is attentive, pure following may match our method, and we expect it to. The proposal only earns its keep in two situations: when hazards appear in regions only the robot's end passes through, and when the leader is inattentive. H2-COMPACT's own design, which delegates all avoidance to the human, is evidence that this case has not been addressed. The experiments below separate the regimes explicitly so a reviewer can see where the benefit comes from, or whether it disappears.

Why not just stop or crawl at every occlusion? Sun et al. show that uniform conservative braking at occluded intersections causes deadlocks, while reading other agents keeps traffic moving. In co-carrying, the file's analysis indicates an unprompted halt produces large resistive forces against a moving leader. Our OA-MPC baseline is essentially this rule made principled. The impulse and deadlock comparison against it tests the claim directly.

Why not just add infrastructure cameras or collaborative perception? This is often simply unavailable in homes, construction sites or stretcher transport. It also degrades under packet loss and misalignment, as discussed around V2VNet and CoBEVT. Where it is available, it is complementary: the gate could treat an external view as another visible region. We do not claim superiority in instrumented environments and will not test there.

Why not just use calibrated continuous intent without discrete modes? Conformal sets over continuous velocities are possible, but they give regions whose safety implications are hard to check against reachable sets in real time. RCIP's result, fewer clarification requests (8% to 58% less than KnowNo) while holding target success, depends on sets that map cleanly to actions. The heuristic-trust baseline, together with a continuous-conformal variant if time permits, tests whether discreteness actually matters.

Proposed Experiments

Offline calibration. We train the mode predictor on the H2-COMPACT dyadic dataset: 96 trials, eight primitives, four payload weights. There a blindfolded follower plays exactly the occluded role, which makes the data a natural fit. Held-out payload weights test whether coverage survives distribution shift. We report empirical miscoverage against $\alpha$ and average set size. If coverage breaks across payloads, we fall back to weight-conditional calibration.

Simulated blind-corner transport. We build co-carrying corridors in Isaac Gym, the simulator used by H2-COMPACT and Skill-Nav, with blind corners, doorways and hidden agents emerging at up to the OA-MPC speed bound. Human leaders are modeled by policies fit to the dyadic data, in three conditions. Attentive leaders react to visible hazards. Distracted leaders fail to react at a controlled rate $\rho$. Hazards appear only on the robot's side, outside the leader's view.

Metrics are hidden-agent collision rate, linear impulse, peak interaction force, completion time, deadlock rate and reference jerk. The decisive checks are three. The collision rate inside $V_t$ should track $\alpha + \rho$ as we vary both. Outside $V_t$ it should stay at OA-MPC's level. Impulse should fall well below OA-MPC's.

Hardware pilot. A small study on a mobile co-carrying platform uses a staged blind corner, with the hazard played by a foam obstacle on a cart. It compares our method against OA-MPC-plus-admittance and pure following on impulse, completion time and NASA Task Load Index, as Tac-Nav used.

Falsification. The proposal fails if the argmax ablation performs as well as the conformal gate across all $\rho$. It also fails if pure following matches us in the robot-side-hazard condition. The first result would mean calibration is unnecessary. The second would mean the visibility gate adds nothing.